TW410273B - Method and apparatus for generating and encoding line spectral square roots - Google Patents

Abstract

A novel and improved method and apparatus for encoding line predictive coding (LPC) data in a speech compression system using line spectral square root values is disclosed. A novel and computationally efficient procedure for determining the set of quantization sensitivities for the line spectral square root values is disclosed, which results in a computationally efficient error measure for use in vector quantization of the line spectral square root values. A novel method of weighting the quantization error is disclosed, which accumulates the quantization error in each line spectral square root value and weights that error by the sensitivity of that line spectral square root value.

Description

A7, B7, 4 A, you are expecting ...], please request a correction page of the written statement (August 8S). 5. Description of the invention (a quantified mechanism. Detailed description of the preferred embodiment. Figure 1 illustrates a conventional device. Determine the LPC coefficients (a (1), a (2) ', a (N)) and those LSP frequencies 0 (1)' ω (2), ... 'ω (N)) to generate from those LPC coefficients And encode LPC filter data. N is the number of filter coefficients in the LPC filter. The speech autocorrelation element 1 calculates a set of autocorrelation values, R (0) to R (N), and s (n) of the following equation (1) is obtained from the frame of the speech sample: R (n) = 2s ( k) s (k + nX (1) k = 1 where L is the number of speech samples in this frame on which the LPC coefficient is calculated. In the specific embodiment of this example, the number of samples in a frame is 160 (L = 160), and the number of LPC filter coefficients is 1 0 (n = 1 0). A linear prediction coefficient (LPC) calculation element 2 calculates this lpc coefficient, a (1) to a (N), and thus Set of autocorrelation values, R (0) to R (N). This Lpc coefficient can be determined by using the Durbin recursion (as discussed in "Digital Processing of Speech Signals", 1978, prrentice_Hall, Robina and Xia This is a rule system described in the following equations (2) to (7): ---------- ----I j I ----I- m 1 ^ 1J-l- ^ J-- -I ml--I (Please read the notes on the back before filling out this page) Ministry of Economic Affairs _ Central Standards Bureau Employee Consumption Cooperation Du Printed Ε (0) =: R ( 〇) ，; iI 1 (2) ki = | R (i) K) R (H) bu E0-1); (3) α ^ ⑴ = ki; 1 (4) called ⑴ = aj (1 · 1 )-kj a j .j (1 -1) for 1 uniform si-i (5) (6) Ε⑴ = (i-kAE. -1); and

The paper size of the table is applicable to the Chinese National Standard (CNS) A4 specification (21QX-297W 410273 A7 ----- -__ B7 V. Description of the invention (1) Background of the invention I. Scope of the invention The invention relates to speech processing. In particular, the The invention is a new and improved method and device for encoding unitary coefficients in a linear prediction based on a speech code system. 11-Description of the Related Art Voice transmission by digital technology has become widespread, Especially the application of long-distance and digital radiotelephones. This method has already generated the key, which minimizes the amount of information transmitted through the channel, and maintains the quality of the t-bone re-reading by the message. "If the voice is simple— Sampling this continuous speech and quantifying each sample transmission independently requires too much data rate of ^ thousand: chips per second (/ bps) to achieve reconstructed speech quality similar to that of a general test-like phone call. However, due to The use of speech analysis, followed by the appropriate code, read input, and re-synthesized on the receiver, can achieve the data rate intentionally printed by the Consumers' Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs ^ ------------ {Please read first Note the surface to fill out this page) extracted by the parameters of the model of the human voice to produce voice compression device for the language% Jin encoder. This device consists of an encoder that decomposes the input speech to extract the associated parameters, and a decoder that uses this to receive in the entire transmission channel .. The decoder then synthesizes the speech. In order to accurately represent the k-time variation of the bone-bone signal, the etiquette of this ceremony department is updated periodically. This is divided into time blocks or analysis messages, during which these parameters are calculated and quantified. These quantized parameters are then passed through the entire transmission channel, and the speech is reconstructed on the receiver from these quantized parameters. This stimulated code linear predictive coding (CELP) method is used in many speech compression specifications -4 paper ruler based on financial standards (CNS) A4 specification (210X297 mm)% 85109891 Chinese manual revision page (August 88) 5 Description of the invention (The N LPC coefficients are called aj (1 0) 'for 1 sj < N. The operation of this element 1 and 2 is well known. In the specific embodiment of this example, the formed waver Is a tenth-order filter, which means that there are 11 autocorrelation values, R (0) to R (1 0), calculated from the autocorrelation element 1, and 10 L p C coefficients, a (1) to a (10) 'is calculated by the LPC calculation element 2. The LS P calculation element 3 converts this set of LPC coefficients into a set of L · S P frequencies with values ω i to N. The operation of this LS P calculation element 3 is well known And described in detail in the aforementioned US Patent No. 5,414,796. For the promotion of the use of this LS PF frequency, given in the paper "Line Spectrum Pair (ls PF) and Speech Data Compression", ICASSP '84' Song By Zhuang Zhuang. The calculation of this LSP parameter is shown in the following equations (g) and (9) together with Table I. This LS frequency is Wherein ν between the root and the following equations billion redundant, ύύ Ρ⑽ Ρν / 2, 1 · 2 + Pn / 2-1 COS ω + ... + Pj cos

'N 1 · ω

丄 N + cos ——ω (8) T (Please read the precautions on the back before filling in this page). Equipment., Ιτ ij— 线 q (co). * 9N / 2 4μ / 2-ι C0S ω + * +% ≪ = os Printed by the Consumer Cooperatives of the Central Bureau of Standards of the Ministry of Economic Affairs

N 1 _ω N -t- cos —ω, (9) where the values of Pn and qn are defined recursively for n = 1, 2, ..., N / 2 in Table I-8-This paper scale applies to China National Standard (CNS) A4 specification (210X297 Gongchu) Employees' Cooperative Cooperative Seal of the Central Standards Bureau of the Ministry of Economic Affairs 410273 at B7 V. Description of the invention (2) An example of the CELP coding rules is described in the paper, a 4.8 kbpS stimulated Code linear predictive encoder,, is proposed by Thomas E. Tremen et al. In the minutes of the 1988 Mobile Satellite Symposium. -An example of this type of 5-way bone encoder that is particularly efficient is detailed in US Patent No. 5,414,796, entitled "Variable Rate Speech Encoder" and assigned to the assignee of the present invention, and herewith Notice. Many speech compression rules use a filter to model the spectral size of this speech signal. Therefore, the coefficients of this filter are calculated using the linear prediction technique for each speech frame. This filter is called linear prediction coding (Lpc ) Filter. Once the filter coefficients have been determined, this filter coefficient must be quantized. An effective method for quantizing this LPC filter coefficient can be used to reduce the number of bits needed to encode this speech signal Element rate. A method for quantizing the coefficients of the LPC filter includes transforming the filter coefficients into line spectrum pair (LSP) parameters and quantizing the LSP parameters. The quantized LSP is then converted back to the LPC filter coefficients, which are Decoder is used for speech synthesis model. Quantization is done in the LSP domain because LSP parameters have better quantization characteristics than LPC parameters, and because of this quantized ls P parameter The sequence characteristic guarantees that the resulting LPC filter will be stable. For a specific set of LSP parameters, the quantization error-difference in a parameter will cause a large change in the response of the LPC filter, and therefore a ratio of Similar quantification errors in another LSP parameter can result in greater perceived degradation. The sensible effects of this quantification can be achieved by allowing more of those LSP parameters that are less sensitive to quantization errors. The quantization error minimizes it. To determine the optimal distribution of this quantization error, it is necessary to determine each LSP parameter. This paper size applies Chinese national standards (CNS > A4 is now (210 > < 2? 7). ） ΫΊ----- In > ί--.---------—— * 丨: | .. W--T. i (#Read the notes on the back before filling this page) Revised Page of Chinese Manual No. 85109891 (August 88) V. Description of Invention (10) 2 ^ i (Nl) zN-NLJ, (N_2) zN-2 + ^ Ji (1) 2 + Jt (〇) 2 ′ 2 · χΓζ + ΐ) ρ (N + 1) ζN + 1 + P (Ν) 2ν +; + ρ (ΐ) ζ + ρ (〇); Yang and the cosine value of the line spectrum with even index (X 2 , X 4, X 6 and so on), long The division is done as follows: • J1 (N-1) zn " 1 + j | (N-2) 2 ^-^ ... +) ^ 1) 2 + 1, (0) z2-2-Xj · ζ + ΐ ) 〇 (Ν + 1) ζN + 1 + Q (N) zn + " .. V + Q (1) z + Q (〇). (2 ·. · If i is odd, Ji (k) = Ji (N + lk). (25) Because symmetry ', only half of this division needs to be done to determine the value of the whole N j 丨. Similarly 'If i is an even number, Ji (k) = -Ji (N + 1 -k), (2 6) (Please read the precautions on the back before filling this page) Θ Install '11 staff of the Central Bureau of Standards, Ministry of Economic Affairs Consumption cooperative printing is opposed to counterclaim, and only half of this division needs to be completed. Sensitivity autocorrelation elements 〇〇6a ~ 06N calculate the autocorrelation of these groups jf, using the following equations:-Ν-π + 1-Rji (n) = Σ Ji (l <)-Ji (k + n) (27) k = 0 -13- National Standard (CNS) A4 Specification (210X297 mm) A7 __B7 410273 V. Description of the Invention (3) Individual Sensitivity of Numbers—for optimally encoding this LSP The preferred method and device for the parameters are described in detail in the commonly pending US Patent Application Serial No. 08 / 286,150, filed on August 4, 1994, entitled "Quantization of Frequency Spectrum Weighted Vector Frequency Quantization", It is assigned to the trustee of the present invention and is hereby notified. SUMMARY OF THE INVENTION The present invention is a novel and improved method and device for quantifying the L pc parameter 'which uses the square root of linear spectrum (LSS)'. The invention converts this LPC chirper coefficients become another set of data, which is easier to quantify than LPC coefficients, and it provides sensitivity to reduce quantization errors, which is the main benefit of LSP frequency coding. In addition, the conversion from LPC coefficients to LSS 値 and Conversion from LS S 値 to LPC coefficient than LPC coefficient and LSP The conversion between related numbers is less computationally intensive. The legend outlines the features, objectives, and advantages of the present invention, which will become more apparent from the detailed description illustrated below. When viewed at the same time as this figure, the same words in it The yuan is the same from beginning to end, where: Printed by the Consumers' Cooperative of the China Standards Bureau of the Ministry of Economic Affairs (please read the precautions on the back before filling in this page). Figure 1 is a block diagram illustrating the LPC coefficients used to generate and encode FIG. 2 illustrates a graph used to redistribute the normalized function of the cosine of the line spectrum in the present invention; FIG. 3 illustrates a method for generating sensitive information. The blockiness of the device encodes the square root of the line spectrum of the present invention; and FIG. 4 is a block diagram illustrating the integer -6 of the square of the line spectrum used to encode this line-this paper is suitable for i-scales; ί! 中 关 家 橾Standard) A4 specification (2 [〇 > < 297 public epidemic > ---- Patent Application No. 85_91 ^ Chinese manual amendment page (August 88) ^ V. Description of the invention (12) Square root of the first-line spectrum The poor first subvector containing △ y Ay 2

AyN (i) ′ is calculated using the subtractor element 115a as:

Ay y (29) △ yi = yi-yi- 丨 < i < N (l) +1 (3 0) The Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs printed this set of values N (1) 'M (2) Etc. 'defines the division of this line spectrum square root vector into the subtractor. In the specific embodiment with the example of N = 丨 〇, the square root vector of the line spectrum is divided into 5 vectors of every two elements, like n (1) = 2, N (2) = 4, and N (3) = ό, N (4) = 8, and N (5) = (l〇). V is defined as the number of sub-vectors. In the specific embodiment of this example, V = 5. In another alternative embodiment, the line frequency square root vector can be divided into different numbers of sub-vectors of different dimensions. For example, in the first sub-vector t is divided into 3 sub-vectors with 3 elements, and three elements of the second sub-vector, and 4 elements in the third sub-vector will be generated, N (2) = 6, and N (3) = 10. In this alternative embodiment, v =; 3. After the first subvector of the square root difference of this line spectrum is calculated in the subtractor i 15a, it is quantized by the elements 116a, 117a, U8a, and 119 &. Element 118a is a codebook of the square root difference vector of the line spectrum. In the specific embodiment of this example, there are 64 such vectors. This line spectrum square root difference vector ^ = codebook can be determined using the well-known vector quantization exercise algorithm. The finger generator 1, element 117a, provides a codebook index, m, to codebook I, 118a. The coding book element 118a should be added to the index to generate the first coding vector. 15- The paper size is applicable to the Chinese national standard (QMS) A4 specification (210X 297 cm). J --- 1 ---- 0 丨 install-- --- Order ----- j 丨 line f, please read the notes on the back before filling in this page} A7 B7 4 A Touch 0 you expect ·] Please request the amendment of the case certificate (August 8S) V. DESCRIPTION OF THE INVENTION (Quantitative mechanism. Detailed description of the preferred embodiment. Figure 1 illustrates a conventional device, by determining the LPC coefficients (a (1), a (2) ', a (N)) and by those LPCs The coefficients generate LSP frequencies 0 (1) 'ω (2), ...' ω (N)) to generate and encode LPC filter data. N is the number of filter coefficients in the LPC filter. The speech autocorrelation element 1 calculates a set of autocorrelation values, R (0) to R (N), and s (n) of the following equation (1) is obtained from the frame of the speech sample: R (n) = 2s ( k) s (k + nX (1) k = 1 where L is the number of speech samples in this frame on which the LPC coefficient is calculated. In the specific embodiment of this example, the number of samples in a frame is 160 (L = 160), and the number of LPC filter coefficients is 1 0 (n = 1 0). A linear prediction coefficient (LPC) calculation element 2 calculates this lpc coefficient, a (1) to a (N), and thus Set of autocorrelation values, R (0) to R (N). This Lpc coefficient can be determined by using the Durbin recursion (as discussed in "Digital Processing of Speech Signals", 1978, prrentice_Hall, Robina and Xia This is a rule system described in the following equations (2) to (7): ---------- ----I j I ----I- m 1 ^ 1J-l- ^ J-- -I ml--I (Please read the notes on the back before filling out this page) Ministry of Economic Affairs _ Central Standards Bureau Employee Consumption Cooperation Du Printed Ε (0) =: R ( 〇) ，; iI 1 (2) ki = | R (i) K) R (H) bu E0-1); (3) α ^ ⑴ = ki; 1 (4) called ⑴ = aj (1 · 1 )-kj a j .j (1 -1) for 1 uniform si-i (5) (6) Ε⑴ = (i-kAE. -1); and

The paper size of the table is applicable to the Chinese National Standard (CNS) A4 specification (21QX—297W% 85109891 Chinese manual revision page (August 88)) 5. Description of the invention (The N LPC coefficients are called aj (1 0) 'for 1 sj < N. The operation of this element 1 and 2 is well known. In the specific embodiment of this example, the resulting wave filter is a tenth-order filter, which means that there are 11 autocorrelation values, R (0 ) To R (1 0), calculated by the autocorrelation element 1 and 10 L p C coefficients, a (1) to a (10) 'calculated by the LPC calculation element 2. LS P calculation element 3 converts this The set of LPC coefficients becomes a set of L · S P frequencies with values ω i to N. The operation of this LS P calculation element 3 is well known and described in detail in the aforementioned US Patent No. 5,414,796. For this LS P The promotion of the use of frequency is given in the paper "Line Spectrum Pair (ls P) and Speech Data Compression", ICASSP '84' Song and Zhuang. The calculation of this LSP parameter is given by equations (g) and (9) and Table I is shown together below. This LS PF frequency is the ν roots between 0 and redundant in the following equation, ύύ Ρ⑽ Ρν / 2, 1 · 2 + Pn / 2-1 COS ω +… + Pj cos

'N 1 · ω

丄 N + cos ——ω (8) T (Please read the precautions on the back before filling in this page). Equipment., Ιτ ij— 线 q (co). * 9N / 2 4μ / 2-ι C0S ω + * +% ≪ = os Printed by the Consumer Cooperatives of the Central Bureau of Standards of the Ministry of Economic Affairs

N 1 _ω N -t- cos —ω, (9) where the values of Pn and qn are defined recursively for n = 1, 2, ..., N / 2 in Table I-8-This paper scale applies to China National Standard (CNS) A4 Specification (210X297) Chu 410273 A7 B7 V. Description of Invention (6

Table IP! =-(A (l) + a (N))-1 p2 = ~ (a (2) + a (Nl)) -Pl · P3 = + a (N-2))-p2 a) = -(a⑴-a (N)) + 1 q2 =-(a (2) -a (Nl)) + =-(a (3) va (N-2)) + qj In Table I, a (l値, ..., a of a (N) is the result of the coefficient of 伖 ratio derived from LPC analysis. If the LPC filter is stable, a characteristic of the LSP frequency is one of the two function roots, that is, the lowest root ω i, which is the lowest root of P (ω), the second lowest root, ω 2, is the lowest root of q (ω), and so on. The odd frequencies of these N frequencies are the roots of P. (ω), and the even frequencies are the roots of q (ω). Solving equations (8) and (9) to obtain the LSP frequency is a computationally heavy task. A major source of computational burden in converting this L P C coefficient to the L S P frequency and back from the L S P frequency to the l p C factor comes from the extensive use of trigonometric functions. One way to reduce the complexity of this calculation is to replace: .HI nn ^ 4 ^^^ 1 HI «I-Φ f poem first ¾ note on the back ¾ then .. fill in this page} ij Central Standard of the Ministry of Economy Printed by Bureau Shellfish Consumer Cooperative

X cos ω (10) 値 for cos (n) of η > 1 can be expressed by a combination of powers of 7 and is used by the recursion of the following triangle equation: _ -9- The size of the paper is applicable to the Chinese national standard ( CNS) A4 specification (2! 〇χ29 ·· mm) A7 B7 V. Description of the invention (7) cos ((n + l) co) = 2 · cos (co) cos (nco)-cos ((η-1 ) ω) · The expansion of this equation can be expressed as (11) cos (2co)-2 · cos (w) ς〇5 (ω)-cos (0) = 2 -1, · cos (3co) = 2 * .cos (cd) cos (2cd) · cos (ω) =. 2 χ (2 χ2 _ι) _χ = 4) < 3 (12) 3x, (13), etc. General power terms, equations (8) and (9) are reduced to polynomials of X: may ---- 0 ^ —I (please read the notes on the back before filling this page) ρ (χ);

Pn / 2 2 Ν + Ρν χ + Ρν χ2 + ... + PP2 + χ2 '-1 (14) q (x):, ^ N / 2 Ν Ν Order ^ ~ + ciN, χ + ^ Ν XZ + ... + qiX2 +: ZZ. '_1 — ~ 2 (15)

Employees' Cooperatives of the Central Bureau of Standards, Ministry of Economic Affairs, I. Therefore, it may generate the information generated by the Lsp frequency (ω 1. _ · Ω Ν) by providing 値 (χ1. · × Ν), which is called the linear spectrum cosine (Xi). Determining the N line spectrum cosine 値 means finding equations (丨 4) and (i 5) ν roots. This procedure does not require the search for triangles, which greatly reduces the computational complexity. The problem of quantizing the cosine of this line spectrum is the opposite of the Ls p frequency. This is the line spectrum cosine close to + 1 and -1. The kidney is very sensitive to quantization noise. 3 In the present invention, the cosine of this line spectrum is useful for quantization. Noise is more robust by converting it into a set of chirps, referred to herein as line frequency. Spectral Square Root (LSS) 値 (y 1 _.. YN). Used to convert the cosine (χ 丨 χΝ) of this line spectrum to the line spectrum-10- This paper applies the Chinese national standard _ (CNS) A4 (210 × 297) according to the standard ^ ---- 410273 A7 B7 V. Invention Note (8) The calculation of the square root (yi ... yN) 値 is shown in the following equation (16). · Yi -Λ / ϊ-χί; 〇 ^ xi < l

1- —yl + xj;-1 < xj < I (16) Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs where Xi is the i-th line spectrum cosine 値 'and yi is the corresponding square-line line root value. The conversion from the line spectrum cosine to the square root of the line spectrum can be regarded as a conversion from the line type spectrum cosine to LSPs ′ ω = arec0s (x), which is approximately proportional. Fig. 2 illustrates a -graph of a function of equation (6). Because of this conversion, the square spectrum root of this line is more sensitive to quantization noise than the line spectrum cosine, and has characteristics similar to the frequency of L s p. However, the conversion between the L PC coefficient and L S S 値 only requires the product of the product and the square root, which is less intensive than the trigonometric calculation required for the conversion between the L PC coefficient and the L SP frequency. In a specific embodiment of the -T kinds of improvements of the present invention, the square root of the line spectrum 値 is coded according to the calculated sensitivity ，, and the code selection method and device are described herein. The method and device for compiling the square root of the linear spectrum of the present invention maximize the presence-perception quality of the coded speech with the smallest number of bits. FIG. 3 illustrates the device of the present invention for generating a line spectrum cosine 値 (χ (丨), χ (2) ', ·, ·, X (N)) and a quantized sensitivity (s 1 's 2'. _., S N). As previously described, N is the number of filter coefficients in the L PC filter. Speech autocorrelation element 1 〇 丨 Calculate a set of autocorrelation 値 'R (0) to r (N), from the speech sampling frame i according to the above equation (1) -------- (Please read the back first Note again. Fill out this page)

Printed by the Consumer Cooperatives of the Central Bureau of Standards of the Ministry of Economic Affairs A 7 --------- B7 V. Invention Description (9) s (n). Linear prediction coefficient (LPC) calculation without the calculation of 102 LPC coefficients, a (1) to a (N) 'from this group of autocorrelation 値, r (0) to r (N), as the above equation # (2) ~ (7). The linear spectrum cosine calculation element ι03 transforms this set of L · PC coefficients into a set of linear spectrum cosine 値, χ 1 to X n, as described in the above formulas (14) to (15). The sensitivity calculation element ι08 generates a sensitivity 値 (Si -_, Sn) as described below. P & Q ° Decimal calculation element 104 calculates two new unitary vectors, p and q, from the lp C coefficient, using the following equation (1 7) 7 (2 2): P (〇) = l, P ( N + 1) = iP⑴ = -a (i)-a (N + li). 〇 < i < N + l Q⑼ = 1 Q (N + 1) = -i Q (i) = -a (i) + a (N + li); 0 < i < N + l Λ1 / Λ) / ΛΪ / \ Jy) /) / z 8 9 ο 1 nx one—ί 1 polynomial division element 105a ~ 1 〇5N complete the polynomial division by Provide these .groups UJU, consisting of Ji (l) to Ji (N), where 'i is the finger 槔 of the cosine 値 of the line spectrum, from which the 値 of the sensitivity is calculated. For the linear spectrum cosine 値 (Xl, x3, x5, etc.) with an odd index, the long division ^ is completed as follows: _ -12- This paper size applies the Chinese National Standard (CNS) A4 specification (210X29 "? mm) ^^ — / 1 It— * l ^ i If— If flftv ^^^^ 1 — '^^ — cv i ^ i ^^^^ 1 ^^^^ 1 ^^^^ 1 11, J, J ( Please read the notes on the back before filling in this page) Revised Chinese Manual No. 85109891 (August 88) V. Description of Invention (10) 2 ^ i (Nl) zN-NLJ, (N_2) zN-2 + ^ Ji (1) 2 + Jt (〇) 2 '2 · χΓζ + ΐ) ρ (Ν + 1) ζN + 1 + P (N) 2ν +; + ρ (ΐ) ζ + ρ (〇); Yang and for Cosine of line spectrum with even index (X 2, X 4, X 6 etc.), long division is done as follows: • J1 (N-1) zn " 1 + j | (N-2) 2 ^-^. .. +) ^ 1) 2 + 1, (0) z2-2-Xj · ζ + ΐ) 〇 (Ν + 1) ζN + 1 + Q (N) zn + " .. V + Q (1) z + Q (〇). (2 ·. · If i is odd, Ji (k) = Ji (N + lk). (25) Because of symmetry, only half of this division needs to be done to determine the whole N Value. Similarly, if i is even, Ji (k) = -Ji (N + 1 -k), (2 6) (Please read the notes on the back before filling this page) Installed in the 11th line of the Consumer Standards Cooperative of the Central Bureau of Standards of the Ministry of Economics because of objections to counterclaims, only half of this division needs to be completed. Sensitivity autocorrelation elements 丨 〇6a ~ 1 06N To calculate the autocorrelation of these groups jf, use the following equation :-: Ν-π + 1-Rji (n) = Σ Ji (l <)-Ji (k + n). (27) k = 0 -13- National Standard (CNS) A4 specification (210X297 mm) A7 B7 410273 V. Description of the invention (11) Sensitivity cross-correlation element 107a ~ 107N Calculate the sensitivity of the square root of the spectrum of this line 値 Use the cross-correlation from speech, R, and RJ i _ group with autocorrelation 値, And weight this result with 1-I Xi I. This operation is done according to the following equation (28): — (H4 Bu ,. 'I ;,' — ㈣. Figure 4 illustrates the device of the present invention, which is used repeatedly. And quantize the set of square roots of the line spectrum. The present invention can be implemented in a digital signal processor (DSP) or programmed in an application specific integrated circuit (ASIC) to perform this function, as described herein. The elements 111, 112, and 113 operate as described above in Figure 3, Fang Cheng 10 1, 102, and 103. The line spectrum cosine calculation element 113 provides this line spectrum cosine 値 (X i, ..., XN) to the line spectrum square root calculation element 12 1 , Which — Calculate the square root of the spectrum of this line, y (1)... Y (N), according to equation (16) above. The sensitivity calculation element 1 Η receives the line spectrum cosine 値 (X i,..., XN) from the line spectrum cosine calculation element 113, and the LPC 値 (a (1), ...) from the LPC calculation element 112 • , A (N)) and the autocorrelation 値 (R (0),..., R (N)) from the speech autocorrelation element 111. The sensitivity calculation element Π4 generates a sensitivity group ，, S1, ..., SN, as described in the description of the sensitivity calculation element 108 in Xiao'3. ~ Once the line square root 値 group of this line is calculated, y (1) ·. .Y (N), and the sensitivity group, S!, ..., SN, the quantization of the line root 値 of the line spectrum is started. -14- This paper size applies to Chinese National Standard (CNS) A4ii (210X 297 mm) J ---! ----- • Installation --- „---„ — Order ------ (谙 read first) But on the back, please fill in this page again.} No. 85_91 patent application printed by the Employees' Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs ^ Chinese manual amendment page (August 88) ^ 5. Description of the invention (12) Square root of the first-line spectrum The poor first subvector containing △ y Ay 2

AyN (i) ′ is calculated using the subtractor element 115a as:

Ay y (29) △ yi = yi-yi- 丨 < i < N (l) +1 (3 0) The Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs printed this set of values N (1) 'M (2) Etc. 'defines the division of this line spectrum square root vector into the subtractor. In the specific embodiment with the example of N = 丨 〇, the square root vector of the line spectrum is divided into 5 vectors of every two elements, like n (1) = 2, N (2) = 4, and N (3) = ό, N (4) = 8, and N (5) = (l〇). V is defined as the number of sub-vectors. In the specific embodiment of this example, V = 5. In another alternative embodiment, the line frequency square root vector can be divided into different numbers of sub-vectors of different dimensions. For example, in the first sub-vector t is divided into 3 sub-vectors with 3 elements, and three elements of the second sub-vector, and 4 elements in the third sub-vector will be generated, N (2) = 6, and N (3) = 10. In this alternative embodiment, v =; 3. After the first subvector of the square root difference of this line spectrum is calculated in the subtractor i 15a, it is quantized by the elements 116a, 117a, U8a, and 119 &. Element 118a is a codebook of the square root difference vector of the line spectrum. In the specific embodiment of this example, there are 64 such vectors. This line spectrum square root difference vector ^ = codebook can be determined using the well-known vector quantization exercise algorithm. The finger generator 1, element 117a, provides a codebook index, m, to codebook I, 118a. The coding book element 118a should be added to the index to generate the first coding vector. 15- The paper size is applicable to the Chinese national standard (QMS) A4 specification (210X 297 cm). J --- 1 ---- 0 丨 install-- --- Order ----- j 丨 line f Please read the notes on the back before filling out this page} Printed by the Employees' Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs 410273 A7 _ B7 V. Description of the invention (13), component △ y ^ m), ..., △ yNp〆!)!). The error calculation and minimization element 116a calculates a sensitivity weighted error, E (m), which represents an approximate spectral deviation, which may incur the original subvector quantizing the square root difference of the line spectrum to this codebook of the square root difference of the mth line spectrum . In the specific embodiment of this example, E (m) is calculated by the following equation: err two 0; (31) E (m) -0; (32) for k = 1 to N (l) (33 Err = err-ι--Δγ ^ (ηα) ·, (34) H (m) = E (m) + err ^ (35) end loop '»(36). E (m) is in the LSS value The sum of the sensitivity weighted squared errors. The procedure used to determine this sensitivity weighted error is described in equations (3 1) to (3 6). The quantization error is accumulated in the square root 値 of each line spectrum and is calculated by LS S 値. Sensitivity weights this error. Once E (m) has been calculated for all coded vectors in the codebook, the error calculation and minimization (ERROR COMP. AND MINI.) Element 116a-selects the index m 'which makes E (m ) Is the smallest. The 値 of this m is this choice: the index '1 of the codebook 1 is called 11. This △ yi' ..., △ y N () quantized 値 is expressed as △ yi… ΔΥν ⑴ and set equal to ΔyTTh), ..., A-yN ^ A). In the summation benefit element 119a, the square root of the quantized line spectrum in the e-th sub-vector is calculated as follows: '-16- This ϋZϊ applies the Chinese National Standard (CNS) A4 specification (2! 〇χ57 公 楚> ίI- , ----- # 装 ---.— I—1Τ ------ Q. (Please note the notes on the back before writing this page) ^ 10273 A7 B7 V. Description of the invention (14? Ι = Σ ^. K = l (37) The square root of the quantized line spectrum 値 yN (1) is calculated in block 119a, and yi for 1 from N (1) + 1 to N (2) is calculated from this line The first sub-direction egg with the square root difference of the spectrum includes ⑴ + ι, ⑴ + 2, ·· △ ΥνΡ) as follows Δ / ΐ = yN (l) + l-yN (l) Δγι = yi * yi-i; · N (l) < i < N (2) +1 (38) (39) f 谙 Read the note on the back first and fill out this page} I use it for the operation of this indicator 値 I 2. Use as follows The same method of selecting 11 is done. This retained sub-vector continues to be quantized in the same way. Operations on all sub-vectors are actually the same, and for example the last sub-vector, the V-th sub-vector, in all The subvectors from 1 to V-1 are quantized and quantized. The Vth of the square root difference of this line spectrum The sub-vector is calculated from the element 115 V as follows.

AyN (V-i) + i = yN (v-i) + i-YN (VT1) · (4〇)

Ayi = Δ / i-AYi.v,: N (yl) < i < N (V) + J. (4J) The Vth subvector is found in the codebook V 第 ·· for E (m ) The smallest coding vector is quantized, which is calculated by the following circuit: -17- This paper size applies the Chinese National Standard (CNS) A4 specification (210X 297 mm)

1T Yinli 1 410273 A7 A7 B7 Central Labor Bureau of Ministry of Economic Affairs Consumer Cooperatives V. Description of Invention (15) err = 0; (42) E (m) = 0; (43) for k = N (V-1 ) +1 to N (V) (44) err = err + Ayk-Aykim) (45) E (m) = E (m) 4 * Sk err ^ (46) end loop '(47) Once this v The optimal coding vector of each sub-vector, the difference between the square root of the quantized line spectrum and the square root of the quantized line spectrum, are calculated for the sub-vector as described above. This process is repeated until all sub-vectors are quantized. In Figures 3 and 4, this block can be completed as a structured block to perform the specified function, or this block can be expressed as a digital signal processor (DSP) or an application specific integrated circuit (ASIC). Functions in the program. The functional description of the present invention allows a person of ordinary skill to complete the present invention in a DSP or ASIC without undue experimentation. Employees' Cooperatives of the Central Bureau of Standards, Ministry of Economic Affairs, Consumer Cooperatives (please read the notes on the back before filling out this page) The previous description of this preferred embodiment is to provide any person with this skill to make or use the invention . The differences between these specific embodiments will be obvious to those skilled in the art; and the general principle of solitary definition can also be applied to other specific embodiments without the permission of this invention. Therefore, the present invention is not intended to be limited to the specific embodiments shown here, but to give the broadest scope having the principles and novel features disclosed herein. -18-This paper size applies to Chinese national standard (CNS > Α4 size (210X 297mm)

Claims (1)

410273 ABCD Printed by the Central Bureau of Standards of the Ministry of Economic Affairs and Consumer Cooperatives 6. Scope of Patent Application 1-A device for generating and encoding linear prediction code (LPC) coefficients in a linear prediction encoder, including: LPC generator A device for receiving digitized speech samples and generating a set of LPC coefficients for the digitized speech samples according to a linear prediction code format; a line spectrum cosine generator device for receiving the set of LPC coefficients and according to a line spectrum The cosine conversion format generates a set of line spectrum cosine 値; and a line spectrum square root device for receiving the set of line spectrum cosine 値 and generates a set of line spectrum square root 値 according to a square root conversion format. 2. The device according to item 1 of the applied patent garden, wherein the square root conversion format is: V1-Xi; 0 ^ xi < 1 where Xi is the i-th line spectrum cosine 値, and 7i is the corresponding i-th Line frequency helps square root. 3. The device according to item 1 of the scope of patent application, further comprising:-a polynomial division device for receiving the set of line spectrum cosine 値 and a set of linear prediction code (LP C) coefficients, and deriving a predetermined The polynomial division format generates a set of quotient coefficients; and T-r- sensitivity cross-correlation device for receiving the set of quotient coefficients, the set of line spectral cosine 値 'and a set of speech autocorrelation coefficients, and according to a weighted cross- 19- It is applicable to Chinese national standards (CNS > M specifications (2I0X297)) on each paper scale; --- • 一 ---- ^ --- Γ -------- ^ line (please read first Note on the back ^; write this page) 2 Yi A8 B8 C8 D8 VI. Calculate the set of square root sensitivity coefficients of a line spectrum based on the patent application range calculation format. 4. According to the third item of the patent application range, the device also includes a sensitivity. The degree autocorrelation device is placed between the polynomial division device and the sensitivity cross-correlation device to receive the set of quotient coefficients' and generate a set of sensitivity autocorrelations to the set of quotient coefficients according to a predetermined autocorrelation calculation format. 5. Installation according to item 3 of the patent application park The device further includes a vector calculation device placed before the polynomial division device to receive the set of LPC coefficients and generate a set of vectors according to a predetermined vector generation format. The vector calculation device calculates the two vectors p and Q in the set of vectors according to the following equation: Please note that P (〇) = 1 P (N + 1) = 1. P (i) =- a (i)-a (N + li) Q (〇) = i Q (N + i) = -i Q (i) = -a (i) + a (N + li);, 0 < i < N + l 0 < i < N + l. Ordered by the Central Ministry of Economic Affairs, the Ministry of Economic Affairs, the Staff Consumer Cooperative Association — 7. The device according to item 6 of the patent application park, where the polynomial division device is set for the odd-numbered line spectrum Square root yield ^ Quotient coefficients of this group J i + Ji (N-2) zN: -2+, + j, (1) z + j; w 2'Xj · 2 + 1 | P (Ν + 1) ζν + 1 + P (M) zn + P (l) z + P (0), line -20- _Applicable to the Ministry of Economic Affairs of China_Central Bureau of Standards printed by employee consumer cooperatives 410 ^ 73 A8 B8 C8 ~ ____ D8 Read patent range », 2 is a polynomial variable, X i is the cosine of the i-th line spectrum 値, and N is 遽The number of connectors of the device. According to the device of the sixth scope of the patent application, the polynomial division device J, ^ 'gen generates the set of quotient coefficients Ji for the square root of the even-line spectrum according to the following equation:' 2 -I -ΙΚΝ-Z) ^-2 ·!-,,,-, ·} ^!) Z + Ji (0) z + 1JQ (N + 1) zn +! + Q (N) zn +.,. Η- Q ( l) z -i-QW :: where 2 is a polynomial variable 'X i is the 乂, line spectrum cosine 値' and N is' the number of wave filter connectors. 9 'The device according to item 3 of the Shenlong Patent Fanyuan, wherein the sensitivity cross-correlation mount generates the square root of the line spectrum according to the following equation: "N' 1 si = (l- | xi |) · R (0) RJi (0) + 2- £ RCk) _; RJi (k),-lc " 1 «where Xi is the square root of the i-th line spectrum 値, R (k) is the k-th bone autocorrelation of this 値 speech sample The coefficients and RJ (k) are the k-th autocorrelation coefficients in the quotient of this group. 10. A system for generating and coding linear prediction code (LPC) coefficients in a linear prediction encoder. Contains: r- An LPC generator with an input to receive digitized speech samples, and an output to generate a set of LPC coefficients; (Please read the precautions on the back before filling out this page) Installation ·, 1T- 21 · This paper standard is applicable to China National Standard (CNS) A4 specification (2l0x297) «0273 Patent application scope ABCD Printed by male workers' consumer cooperative of Central Standard Bureau of Ministry of Economic Affairs with one round of line cosine generator connected to An output of the LPC generator; and a line spectrum square root generator having an input connected to the line frequency Cosine generator output, and having an output. U-root, the first system of the scope of patent application, further includes: less-to-type divider calculator 'has — input connected to the round-out of the line spectrum square root generator' and has an output; and a sensitivity cross-correlation calculator , Has an input connected to the output of the term division calculator, and has an output & 12 'according to the H patent of the patent application of the nth term of the 1-inclusive-sensitivity self-phase calculator placed in the polynomial division The difference between the calculator and the desensitization cross-correlation calculation is that there is an input connected to the output of the polynomial division calculator and an input 0 U connected to the sensitivity cross-correlation calculator. The method for generating and encoding linear prediction code (LPC) coefficients in the processor 'includes the following steps: generating a set of LPC coefficients for the digitized speech sample according to the linear prediction code format; and generating _group: line spectrum according to the first-line spectrum cosine transformation The cosine chirp and a set of linear spectrum square root chirps are generated according to a square root transformation format. 14. The method according to item 13 of the scope of patent application, in which the step of generating the square root of the set-line spectrum includes: -22- This paper size applies the Chinese National Standard (CNS) Α4 specification (2 "0 × 297 公 & (Please Please read the notes on the back before filling this page), 1T 410273 A8 B8 C8 D8 patent application scope yi 2; v 0 < XJ < 1 1-2 ^ / 1 + xi '· 1 <x; < 0 where xi is the cosine 値 of the i-th line spectrum and sum is the square root 値 of the i-th line spectrum.! 5. The method according to item 13 of the patent application scope further includes the following steps: Division according to a predetermined polynomial Format, generating a set of quotient coefficients; and calculating a square root sensitivity coefficient of a group line spectrum according to a weighted cross-correlation calculation 袴 _ _ formula. According to the method of _Please Patent Fanyuan Item 15, it further includes The self-phase_calculation format generates a set of sensitivity autocorrelation and gives the set of quotient coefficients. + 17 · The method according to item 15 of the scope of patent application, further including the step of generating a set of vectors according to a predetermined inward generation format. 0 18.According to item 1 of the scope of patent application Method, in which the step of generating a set of vectors includes the following steps: ——. ·! 装 —— (锖 Please read the notes on the back and fill in this page) Order the staff consumer cooperative of the Central Standards Bureau of the Ministry of Economic Affairs to print P (0 ) = 1 P (N + 1) = 1 P (i) = -a (i)-a (N + li) Q (〇) = 1 Q (N-fl) =-] Q («) = -a (i) + a (N4-li); 23 paper rule ^^ (2Ι0 × 297 | 0 < i < N 弋 1 0 < i < N + l. ABCD 410273 'Scope of patent application 19' According to the scope of patent application No. 18 Term method, in which the step of generating a set of quotient coefficients J i to the square root of the odd-line spectrum 骒 includes completing the following polynomial division: Ji (Nl) zN ~ 1 + J; (N-2) zN ~ Z + ... -fJ; (l) 2 + Ji (0) ζ2-2 · χ; -2 + ι) ρ (Ν + 1) ζν + 1 + Ρ (Ν) ζ ^ + + Ρ (1) ζ + P (〇 ), Where ζ is a polynomial variable, Xi is the cosine of the i-th line spectrum 値, and N is the number of wave device connectors. 2__ According to the support of item 18 of the scope of the patent application), a set of quotient coefficients is generated The step of giving the square root of the spectrum of the even line by J i includes completing the following polynomial division: I, t + ί | ίΟ) 2 π- Q (N) ^ H. Q (1) z + C ^ ;, Where Z is a polynomial variable, Xl is the "fixed-line spectrum cosine 値", and N is the number of filter connectors. J.IJ ·! Equipment I— (Please read the precautions on the reverse side before filling out this page) Printed by the Staff Consumer Cooperative of the Central Bureau of Standards of the Ministry of Economic Affairs -24 < 297 mm)